While Large Language Models (LLMs) have made remarkable progress over the past few years and are already changing how people work, they remain costly, energy-intensive, and not always reliable, especially when outputs are generated from partially unvetted or low-quality data. At Covalo, we strongly believe that the most impactful approach for the beauty and personal care industry is not to rely solely on general-purpose models, but to combine narrow, task-specific AI models with high-quality domain data and experts in the loop.

By training models on well-defined tasks, grounding them in vetted sources, and keeping domain experts involved to supervise, validate, and iteratively refine outputs, it is possible to deliver strong results while optimizing for accuracy, robustness, and energy efficiency.

A recurring theme in discussions with industry professionals is the lack of time to focus on truly meaningful work. A significant share of R&D, regulatory, and commercial teams’ time is still spent on operational tasks: collecting information, manually entering data, reviewing documents, and navigating fragmented sources. In our view, this is where AI should be deployed first, automating low-value and time-consuming work to free up time and resources for more creative, high-impact activities. Done well, this not only accelerates innovation, but also supports a more engaged and fulfilled workforce.

A concrete example comes from our work on a recommender system for sustainable ingredient alternatives (1), developed jointly with Switzerland’s two leading research institutes, ETH Zürich and EPFL. As part of that effort, we built sophisticated knowledge graphs, taxonomies, and an ontology based on data extracted from more than 50,000 technical documents available on our platform. This work laid the foundation for what we call a semantic layer, which enables highly efficient AI models to perform well on specific, high-value tasks.

One of the most tangible applications is document automation. We built an OCR (Optical Character Recognition) and LLM-assisted pipeline that extracts key information from technical documentation (such as TDS, MSDS, CoA, and regulatory statements) and populates standardized ingredient records. On average, this saves around five minutes of manual work per product. At the scale of the 100,000 ingredients listed on Covalo, this represents more than 8,000 hours of manual effort avoided. Beyond time savings, this approach makes it significantly easier to keep information up to date and continuously improve the underlying ontology and data model.

This, in turn, improves ingredient discovery. Enterprise users can search across both structured data (standardized fields) and unstructured data (documents), and compare information that has historically been scattered across PDFs, emails, and supplier portals. We are applying a similar logic to analytics: enabling users to ask questions in natural language and instantly generate structured reports and visual outputs, work that traditionally required hours of manual effort and specialist skills. For many organizations, the barrier was never access to data, but rather the time and resources required to operationalize it. AI now makes it possible to make more data-driven decisions with far less friction.

Ultimately, the key behind reliable AI in this space is not the model itself. It is high-quality, well-structured data, combined with strong domain expertise to validate outputs, correct edge cases, and continuously improve the system.

In formulation and R&D, the potential of AI is real, and the capabilities already exist. However, the main constraint today is the lack of deep, comparable, and reliable datasets. This is precisely why, during the development of our recommender system, we decided to temporarily shift focus toward data collection, cleansing, normalization, and harmonization. Without that foundation, even the most advanced models will underperform.

Over time, the industry will benefit from additional sources of training data, including high-throughput experimentation and more systematic capture of negative results (failed or suboptimal trials), which are often discarded but are critical for building robust predictive models.

In regulatory and compliance, AI already delivers meaningful value and has enormous future potential. That said, AI is not yet capable of replacing experienced regulatory experts, particularly when it comes to interpretation, contextual judgment, and anticipating regulatory trajectories. Experts still play a critical role in identifying implicit risks (for example, when a substance may be present as an impurity or derivative even if it is not explicitly stated) and in building resilient product pipelines by forecasting which restrictions are likely to emerge.

Even when AI provides significant productivity gains, it remains essential to maintain critical thinking, restrict models to trusted data sources, and systematically challenge and validate results, potentially using AI as a second layer of verification.

One of the biggest emerging risks is that an increasing share of digital content is now AI-generated, and often references other AI-generated sources. As a result, returning to primary sources and ensuring information validity will become increasingly difficult, creating a risk of a misinformation loop in technical and regulatory contexts. This is why digital traceability and provenance, the ability to track where a statement comes from and which source supports it, should become a major focus for industry standards and regulators.

In terms of risks, limitations, and how to mitigate them, the recommendations are simple:

1. Start with data. No data, no AI.
   
   
2. Don’t boil the ocean. The best results come from narrow scope, clear use cases, and disciplined execution.
   
   
3. Don’t do it alone. Combine internal domain expertise with technology specialists, collaborate with peers and customers, accept imperfections, and iterate quickly in a structured way.